fkluger/cuboids_revisited

Robust Shape Fitting for 3D Scene Abstraction

cuboids_revisited's Issues

A few questions

Really cool work! I'm going through the paper and the code and I had a few questions:

1. Is there an easy way to run inference on an individual image? It looks like the answer is no, but it shouldn't too hard for me to patch something together using evaluate.py as a guide. I figure'd I'd check before going through the effort, though.

2. I'm having a bit of trouble understanding the output format. After running evaluate.py, I get a bunch of pkl files in results/eval/test/gt//original/m6_p1_s4096_k1_ic10_rs1. Are these predictions or ground truth (unsure because of the 'gt' in the path)?

3. When looking at a 9-tuple, like the ones found in the 'best_models' entry of the pkl files like this:

[ 0.8778091 , 1.0231334 , 0.01 , -1.4559864 , -1.8047982 , -0.76489663, -0.14193119, -0.21265107, 1.7678456 ]

My understanding is that the first three values are extents in x, y and z and the last three values are the translation (so effectively the center coordinates?). But I'm not entirely sure what to make of the middle three values - I think they're the rotation, but I was naively expecting four values for angle-axis format. I'd like to transform these into a rotation matrix, 3D geometry is not my strong suit, so I suspect I'm missing something obvious.

Thanks for any help you can provide!